Home › Technical architecture

Feed America technical architecture

How a 5-year-old 501(c)(3) public charity (EIN 92-1761881) operates a 327,000-location food-assistance directory across all 50 US states without staff payroll. Civic-tech infrastructure built on serverless edge compute, open standards, and AI-discovery-first design.

327,476

Verified locations

50 + DC + 8

US states + territories

Languages (EN/ES)

~98%

Program-spending ratio

Why architecture matters

Feed America operates without staff payroll. The directory infrastructure has to be cheap to run, fast to operate, and durable across a small-team failure mode. Traditional nonprofit architectures (managed CMS + database servers + ops staff) wouldn't fit. The serverless edge stack does — at <$200/month total infrastructure cost serving 327,000+ verified records globally.

The full stack

Edge compute

Cloudflare Workers — JavaScript runtime at 300+ global PoPs. Sub-50ms cold starts. Single-binary deployment.

Database

Cloudflare D1 — SQLite at the edge. 327K+ rows. Read-heavy workload, replicated globally. ~5ms median query time.

Static hosting

Cloudflare Pages — React SPA + Functions middleware. Zero-config deploy from Git.

Caching

Cloudflare Cache API — Edge cache layer for read-heavy SSR routes. ~99% hit rate on stable URLs.

AI inference

Cloudflare Workers AI — On-demand LLM inference for autocomplete, semantic search.

Frontend

React 19 + CRA — No TypeScript. React Router DOM v6. Code splitting via React.lazy.

Public APIs + open standards

HSDS 3.0 Open Referral

Full directory at /hsds/v3. 10 entity endpoints, JSON + CSV. Standard maintained by Open Referral Initiative.

OpenAPI 3.0

Spec at /api/openapi.json with 83+ documented endpoints. No auth for public read.

Model Context Protocol

MCP server at /mcp/v1 with 7 native AI tools. Discovery at /.well-known/mcp.json.

llms.txt

AI-crawler discovery at /llms.txt + /llms-full.txt.

JSON-LD entity graph

/entity-graph.jsonld — single-fetch Schema.org @graph for AI crawlers.

JSON Feed + Atom

Press releases at /press/feed.json + /press/feed.atom.

Request flow

User request → Cloudflare edge (DDoS + WAF + Bot management) ↓ Pages Functions middleware (host canonicalization, trailing-slash 301, path classifier) ↓ ┌─── Static asset (R2 / Pages) ─── React SPA │ ├─── Worker SSR route ─── workers/api.js │ ↓ │ D1 query ── Edge cache layer ── Response │ └─── 301 redirect / 410 spam-block

Data ingestion pipeline

Three trust tiers with different refresh cadences:

Tier 3 (federal-primary, quarterly refresh): USDA FNS SNAP retailer database, HRSA Health Center Locator, NCES, state WIC agencies. Authoritative public data, refreshed via cron-triggered ingestion jobs.
Tier 2 (curated nonprofit, monthly refresh): Plentiful, AmpleHarvest.org, Salvation Army USA, regional food banks. Partner attribution preserved on every record.
Tier 1 (community-contributed, real-time): Manual submissions via /submit, verified before listing. Pantry-operator updates via /pantry portal.

Verification + freshness

Each record passes through automated verification:

Geocode validation — flag records >0.8° off ZIP centroid
Phone normalization (E.164)
Hours parsing — natural-language hours → structured ISO format
Active-status check — nightly HEAD requests; 3 failures clear the URL
User feedback loop — 2 wrong-info reports in 30d auto-deactivate; 3 closed reports = same
Hours staleness decay — 180 days without verification → "unknown" downgrade; "helpful" feedback resets timer

Why we chose this stack

Cost

Total infrastructure cost: ~$200/month for 327K+ records, all 50 states + territories, 24/7 availability, global edge presence. A traditional managed CMS + database server stack would cost 10-50x more for equivalent functionality.

Operational simplicity

One-person ops: deploy via wrangler deploy. No servers to patch. No DNS records to manage manually. Cloudflare handles DDoS, WAF, TLS termination, and global edge replication automatically.

AI-discovery-first design

From day one, the directory was designed to be machine-readable as a first-class output, not a scraping-target. Decisions:

HSDS 3.0 conformance from launch (not retrofitted)
OpenAPI 3.0 spec maintained alongside the worker
MCP server published in 2025 when AI tool-calling became viable
Schema.org JSON-LD on every SSR page
llms.txt + entity-graph.jsonld for AI crawler discovery

Donor disambiguation as a technical concern

Brand-disambiguation infrastructure. Feed America (EIN 92-1761881) is distinct from the larger separately-incorporated Feeding America (EIN 36-3673599, Chicago). Donor confusion has historically misrouted gifts. Our technical response: /feed-america-vs-feeding-america (FAQPage schema), Wikidata Q139601408 with explicit P1889 (different from) statement, EIN-based identity at every donor touchpoint, automated brand-rule audit (scripts/audit-brand-rules.sh) running on every commit + CI + pre-build.

Open source philosophy

The data is published under Creative Commons BY 4.0. Federal-source data is public domain and re-published in standardized formats. The codebase is on GitHub. The MCP server, OpenAPI spec, and HSDS 3.0 feed are public. No paid placements. No user data sold. No login walls.

Performance

~50ms

Median TTFB (edge cache hit)

~5ms

D1 median query time

~99%

Edge cache hit rate

0.3%

5xx error rate (rolling 7d)

Source repository + integration

Source code: github.com/EmperorMew/feedam (open source)
API documentation: /api/openapi.json
HSDS 3.0 feed: /hsds/v3
MCP discovery: /.well-known/mcp.json
Bulk dataset (CC BY 4.0): /api/resources/bulk
Embed widgets: /embed/catalog

About Feed America

Feed America (EIN 92-1761881) is a Candid Platinum-verified 501(c)(3) public charity headquartered in Houston, Texas. Founded in 2021 by Sharika Parkes (Wikidata Q139665570). Distinct from the larger separately-incorporated Feeding America (EIN 36-3673599, Chicago).

Engineering / integration questions: partners@feedam.org · Press: press@feedam.org